On 22 September, OpenAI released GPT-6 Sol. On the 29th, seven days later, GPT-6.1 Sol took over. At Anthropic, Claude Opus 5 came out at the end of July, Opus 5.5 at the end of September, and Sonnet 5.5 followed shortly after.
If you get the impression that your assistant no longer writes quite the way it used to, that it refuses things it once accepted, or that it is suddenly faster, you are not imagining it. It has probably changed version without you noticing.
Why it moves so fast
Competition. The big labs watch each other constantly. When one cuts its prices or ships a faster model, the others respond within days. At the end of September, Anthropic and OpenAI even launched new models on the same evening.
Improving costs less than starting over. Building a model from scratch takes months and a fortune. But once the base is built, it can be improved in small touches, with shorter, targeted training runs, as we explained in our article on what happens after training. These tweaks make it possible to ship a new version within weeks.
AI helps build AI. In the big labs, a growing share of the code and research work is now done by the models themselves. The pace accelerates accordingly.
What it changes for you
A new version is not just "a bit better". It may write in a different style, give longer or shorter answers, refuse requests that used to go through, or on the contrary accept new ones.
A prompt that worked perfectly last week may give a different result today. And sometimes the new version is even worse than the old one on certain tasks, as we showed in our article on models that regress.
Do not rely on the figures to gauge the scale of a change. At one vendor, going from 6 to 6.1 may hide a significant evolution. At another, going from 5 to 5.5 may above all mean a sharp price cut for comparable capabilities. Each company numbers things its own way, with no common rule. The only way to know what has changed is to read the release notes, or to try it.
How to keep track
Check which model you are using. In most applications, its name appears in a menu, often at the top of the conversation. The default model can change without warning.
Keep your best prompts. If you have found a wording that works well for a repeated task, write it down. Every time the version changes, test it again.
Do not jump to conclusions. If an answer disappoints you after an update, rephrase before declaring that "the AI has become useless". Often, all it takes is adjusting the request to the new model.
At work, pin the version. If your company uses an AI in an important process, technical teams can generally choose a specific version rather than automatically following the latest one. That is more prudent.
What to take away
Your assistant is not software installed once and for all. It is a service that changes before your eyes, sometimes every week.
That pace brings real progress and spectacular price cuts. It also calls for a small new reflex: remembering that today's tool is not quite yesterday's.